Fully Automated Approach for Diagnosis of Supraspinatus Tendon Tear on Shoulder MRI by Using Deep Learning

Rationale and objectives: To explore the feasibility of applying deep learning (DL) approach to detect supraspinatus tendon (ST) tear on shoulder MRI by using arthroscopy as the reference standard.

Materials and methods: In this retrospective study, a total of 431 participants from two different manufacturers and two centers was used to build and validate a DL-based ST tear detection system. The proposed system was developed by using U-Net networks to segment and isolate ST followed by a swin transformer architecture to determine the presence or absence of a ST tear. The Densnet101 and Resnet50 as classifiers were also evaluated. Three radiologists performed subjective diagnoses to obtain the diagnosis results. Receptor operating characteristics (ROC) curves, area under the curve (AUC), sensitivity, specificity, and accuracy were used to evaluate the diagnosis performance of the proposed system and radiologists. We also compared the model's performance to that of radiologists.

Results: The proposed ST tear detection system with the classification network adapted from swin transformer achieved an AUC of 0.986, an accuracy of 0.929, a sensitivity of 0.918, and a specificity of 0.940 for detecting a ST tear, indicating high overall diagnostic accuracy. No statistically significant disparities in diagnostic efficacy were observed between the proposed ST tear detection system and musculoskeletal radiologist.

Conclusion: The proposed ST tear detection system exhibits a diagnostic performance on par with that of seasoned clinical radiologist.